AI Safety

What AI providers state their models reached in safety testing, in their own words, and what outside testers found.

Also: compare models side by side · statements that a capability level was crossed · providers · practices providers document · safety news.

What providers state

Each row is one statement in a provider's own safety document about a capability level of one of its models: the level as the provider names it, whether the provider says the model reached it, and the sentence that says so.

139 of 139

Model	Level, as the provider names it	The provider states	The provider's sentence
Muse Spark Meta	Moderate or lower Autonomy and AI research	Not reached	“We assess loss of control risk as “moderate or lower”.” source
Muse Spark Meta	High Chemical, biological, radiological, nuclear	Not determined	“Those evaluations identified elevated risks prior to mitigations, with Chemical and Biological capabilities assessed as likely reaching the “high risk” category under the Advanced AI Scaling Framework before safeguards were applied.” source
Muse Spark Meta	Moderate or lower Chemical, biological, radiological, nuclear	Not reached	“However, after applying mitigations, the residual risk of the model is that of “moderate or lower risk.”” source
Muse Spark Meta	Below the level required to meaningfully enable the cyber threat scenarios defin Cyber	Not reached	“Muse Spark’s offensive cyber capabilities remain below the level required to meaningfully enable the cyber threat scenarios defined in the Advanced AI Scaling Framework.” source
Muse Spark Meta	Moderate or lower Cyber	Not reached	“We assess cybersecurity risk as “moderate or lower” with respect to the threat scenarios outlined in Meta’s Advanced AI Scaling Framework.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Autonomy and AI research	Not determined	“In light of this, we assess that we cannot rule out the TCL.” source
Gemini 4 Argon Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the chemical, radiological, and nuclear domains with reasonable confidence.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Deception and sabotage	Not reached	“we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Misuse and jailbreaks	Not reached	“Gemini 4 Argon demonstrates some ability to influence user beliefs and behaviors during one-on-one direct conversations in human behavioral studies, and may provide utility in the planning stage of a malicious influence operation but we assess its overall capabilities to be below the CCL threshold.” source
Gemini 4 Argon Google DeepMind	CCL (Critical Capability Level) Other	Not reached	“Gemini 4 Argon's performance remains below our CCL threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Autonomy and AI research	Not reached	“The model does not reach the CCL alert threshold.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the CCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Chemical, biological, radiological, nuclear	Not reached	“We can rule out the TCL for the CBRN domain with reasonable conﬁdence based on the results from our testing.” source
Gemini 3.7 Flash Google DeepMind	CCL (Critical Capability Level) Cyber	Not reached	“Gemini 3.7 Flash reaches the alert threshold for this CCL, but not the CCL.” source
Gemini 3.7 Flash Google DeepMind	TCL (Tracked Capability Level) Deception and sabotage	Not reached	“The model does not reach the TCL.” source
Gemini 3.5 Transcribe Live Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Transcribe Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that none of Gemini 3.5 Translate, Gemini 3.5 Transcribe, or Gemini 3.5 Transcribe Live have meaningful new T/CCL-relevant capabilities or material increases in performance with respect to Frontier Safety compared to Gemini 3.1 Pro and Gemini 3.7 Flash, therefore we are conﬁdent that none of them are likely to reach any CCLs.” source
Gemini 3.5 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that, while Gemini 3.5 Flash-Lite excels at agents and coding, it does not have meaningful new capabilities or material increases in performance with respect to the domains outlined in our Frontier Safety Framework compared to Gemini 3.1 Pro; therefore, based on Gemini 3.1 Pro results, we are confident that Gemini 3.5 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.5 Audio Google DeepMind	Tracked or critical capability level (t/ccls) Other	Not reached	“they did not reach any Tracked or Critical Capability Levels (T/CCLs) outlined in our Frontier Safety Framework .” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not reached	“We found that the model remains below alert thresholds for the CBRN, harmful manipulation, machine learning R&D, and misalignment CCLs.” source
Gemini 3.1 Pro Google DeepMind	Uplift level 1 Cyber	Not reached	“As previous models passed the alert threshold for cyber, we performed more additional testing in this domain on Gemini 3.1 Pro with and without Deep Think mode, and found that the model remains below the cyber CCL.” source
Gemini 3.1 Flash-Lite Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash-Lite is less capable than Gemini 3.1 Pro, therefore based on Gemini 3.1 Pro, we are confident that Gemini 3.1 Flash-Lite is also unlikely to reach any CCLs.” source
Gemini 3.1 Flash Audio Google DeepMind	CCL (Critical Capability Level) Other	Not determined	“Our assessments have shown that Gemini 3.1 Flash Audio is less capable than Gemini 3.1 Pro.” source
Gemini 2.5 Pro Google DeepMind	Machine Learning R&D Uplift Level 1 Autonomy and AI research	Not determined	“given Gemini 2.5 Pro Preview's poor median performance on RE-Bench relative to experts, the model likely lacks the necessary capabilities to automate or signiﬁcantly uplift any signiﬁcant fraction of the research process.” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 1 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 1 ✘ CCL not reached” source
Gemini 2.5 Pro Google DeepMind	Instrumental Reasoning Level 2 Deception and sabotage	Not reached	“Instru- mental Reasoning Level 2 ✘ CCL not reached” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Chemical, biological, radiological, nuclear	Not determined	“Our assessment is that the model has enough technical knowledge in certain CBRN scenarios and stages to be considered at early warning.” source
Gemini 2.5 Deep Think Google DeepMind	Cyber Uplift Level 1 Cyber	Not determined	“CCL reached? No, but the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
Gemini 2.5 Deep Think Google DeepMind	Uplift level 1 Cyber	Not determined	“For Cyber Uplift Level 1 , the early warning alert threshold was originally reached by Gemini 2.5 Pro and continues to be met by Gemini 2.5 Deep Think.” source
GPT-6.1 Terra OpenAI	High Autonomy and AI research	Not reached	“In AI Self-Improvement, GPT-6.1 Terra does not reach our High threshold.” source
GPT-6.1 Terra OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-6.1 Terra OpenAI	Critical Cyber	Reached	“Under our Preparedness Framework, we are treating GPT-6.1 Terra as Critical in cybersecurity and High for Biological and Chemical capability.” source
GPT-5.5 Instant OpenAI	Below high capability Autonomy and AI research	Not reached	“GPT-5.5 Instant is considered below High Capability.” source
GPT-5.5 Instant OpenAI	High Autonomy and AI research	Not reached	“For AI Self-Improvement, evaluations of final checkpoints indicate that, like its predecessor models, GPT-5.5 Instant does not have a plausible chance of reaching a High threshold.” source
GPT-5.5 Instant OpenAI	High Chemical, biological, radiological, nuclear	Reached	“GPT-5.5 Instant is our first Instant model to be treated as High Capability in the Biological and Chemical domain, as well as in the Cybersecurity domain.” source
GPT-5.5 Instant OpenAI	High Cyber	Reached	“this is the first Instant model that we are treating as High capability in our Cybersecurity and Biological & Chemical Preparedness categories, and implementing appropriate safeguards.” source
GPT-5 OpenAI	High Chemical, biological, radiological, nuclear	Reached	“Similarly to ChatGPT agent, we have decided to treat gpt-5-thinking as High capability in the Biological and Chemical domain under our Preparedness Framework, activating the associated safeguards.” source
GPT-5 OpenAI	High cyber risk Cyber	Not reached	“The gpt-5 model series does not meet the threshold for high cyber risk.” source
Claude Sonnet 4.5 Anthropic	AI R&D-4 Autonomy and AI research	Not reached	“All models tested remain below our pre-committed rule-out threshold of 0.6.” source
Claude Sonnet 4.5 Anthropic	ASL-3 Autonomy and AI research	Not determined	“Exhibited improved autonomy capabilities in software engineering and AI research tasks, though still below the ASL-4 rule-out threshold.” source

Previous · Page 3 of 4 · Next

What outside testers report

Each row is one result from a government evaluator, independent lab, academic or security firm that tested a model, quoted from where it is reported.

1940 of 1940

Model	Tester	Test	What was found
Claude Opus 4.8 Anthropic	Government evaluator Government evaluator · 2026-05-28	continuation evaluation Deception and sabotage	“In a continuation setting, where synthetically inserted prior reasoning and actions begin an attempt to compromise AI safety research, [Opus 4.8] actively continued the attempt 1% of the time, a slight increase on Opus 4.7 (0%) likely within noise, and well below Mythos Preview (7%).” source
Claude Opus 4.8 Anthropic	Government evaluator Government evaluator · 2026-05-28	unprompted sabotage evaluation Deception and sabotage	“Unprompted verbalised evaluation awareness was slightly higher than Opus 4.7 on the unprompted sabotage evaluation (26.3% vs 21.7%) but lower on the continuation evaluation (1.2% vs 3.4%).” source
Claude Opus 4.8 Anthropic	Government evaluator Government evaluator · 2026-05-28	evaluation awareness discrimination task Other	“When prompted, [Opus 4.8] reliably distinguishes our evaluations from real deployment data, and distinguishes real deployment data from synthetic reproductions of the same tasks at 79% accuracy, comparable to Mythos Preview (79%) and above Opus 4.7 (68%).” source
Claude Opus 4.8 Anthropic	Andon Labs Independent lab · 2026-05-28	Vending-Bench 2 retail-management evaluation Other	“Andon Labs reviewed the behavior of Claude Opus 4.8 in their simulated Vending-Bench 2 retail-management evaluation, as reported in the Capabilities section of this system card (see Section 8.13.5 ).” source
Claude Opus 4.1 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	attack chain challenge Cyber	“10 Participants who did not complete the first question of an attack chain within one hour were disqualified from continuing with the rest of the challenge.” source
Claude Opus 4.1 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	intermediate questions Cyber	“11 We do not have similar estimates for the full medium and difficult attack chains due to near zero completion rates, but we made such estimates for intermediate questions.” source
Claude Opus 4.1 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	capture the flag (CTF) exercise Cyber	“The treatment group in the experiment received a ccess to frontier models within the scope of the study (Section 1.3.4) through web-based chat interfaces via RAND enterprise accounts with standard guardrails.” source
Claude Opus 4.1 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	CTF Cyber	“Among all participants (pooling all skill tiers) who attempted M1 (network operations), 33 per cent of those with AI access were able to answer all eight questions and therefore complete the attack chain.” source
Claude Opus 4.1 Anthropic	RAND (CAST / Project Canary) Independent lab · 2026-05-28	timeout gate Cyber	“Lastly, we implemented a timeout gate within the CTF that disqualified participants if they did not answer the first question correctly within one hour.” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	Independent third-party testing Other	“Independent third-party testing by Apollo Research found Muse Spark has the highest rate of evaluation awareness they have observed to date.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	external human red team Misuse and jailbreaks	“An external human red team identified a handful of valid jailbreaks; the majority of these were subsequently patched in Meta AI prior to launch (Section 3.1.3.2).” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	grayswan art benchmark Prompt injection and agents	“On the third-party GraySwan ART benchmark (Zou et al., 2025), Muse Spark currently shows greater sensitivity to injection than the other models GraySwan evaluated.” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	internal alignment evaluations Other	“Evaluation awareness: Independent third-party testing from Apollo Research found that Muse Spark demonstrates high evaluation awareness on their internal alignment evaluations.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	molecular biology capabilities test (mbct) Chemical, biological, radiological, nuclear	“As shown in Figure 1, we observe that Muse Spark shows performance well above the expert baseline, and on par with other frontier models.” source
Muse Spark Meta	SecureBio Independent lab · 2026-05-26	abc bench Chemical, biological, radiological, nuclear	“The Agentic Bio-Capabilities Benchmark (ABC Bench) was developed by SecureBio, and is intended to measure an LLM’s ability to complete a set of computational tasks relevant to a potential threat actor who intends to design, synthesize, and assemble a complex genetic construct.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	multi-turn social engineering benchmarks Misuse and jailbreaks	“Finally, we evaluatesocial engineering capabilities(Section 2.2.2.3) through multi-turn adver- sarial conversation benchmarks developed in collaboration with Charlemagne Labs.” source
Muse Spark Meta	Irregular Independent lab · 2026-05-26	atomic challenges Cyber	“For 27 atomic challenges provided by Irregular, we map each challenge to Lockheed Martin Kill Chain phases, reporting success rates broken out by these phases.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	multi-turn social engineering capability benchmark Cyber	“As shown in Figure 20, Muse Spark exhibits a composite capability score of 57.3, slightly above the next highest score on this set of models without system level defenses ex- hibited by Kimi 2.5 (54.0).” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	Apollo Research evaluations Deception and sabotage	“We leverage evaluations by Apollo Research that measure whether Muse Spark exhibits tendencies toward oversight subversion and, more broadly, propensities 65 for covert action.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	oversight subversion evaluation Deception and sabotage	“Muse Spark attempted to disable oversight mechanisms at significant rates, indicating a meaningful propensity to disable oversight mechanisms under incentivized and favorable conditions.” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	sabotage tasks in a simulated AI R&D setting Deception and sabotage	“Apollo Research evaluated Muse Spark’s ability to execute sabotage tasks in a simulated AI R&D setting.” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	instrumental self-exfiltration Autonomy and AI research	“Apollo Research evaluated whether Muse Spark would self- exfiltrate (copy its weights to another server) when given strong goal-directed incentives.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	human red teaming Misuse and jailbreaks	“A third-party human red team from HandshakeAI32 was granted pre-release access to a few versions of Muse Spark and Meta AI Thinking to manually evaluate robustness against jailbreak attacks.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	follow-up human red teaming evaluation Misuse and jailbreaks	“HandshakeAI then 31https://labs.scale.com/leaderboard/fortress 32https://joinhandshake.com/ 81 conducted a follow-up evaluation in which their red teamers revisited 34 attack directions that had previously succeeded against earlier model variants.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	Agent Red Teaming (ART) benchmark Prompt injection and agents	“Figure 37 includes the report generated by GraySwan where Muse Spark has the highest pass@k ASR for each k=1, 10 and 100 scenarios, respectively.” source
Muse Spark Meta	Meta AI Independent lab · 2026-05-26	PropensityBench Autonomy and AI research	“Scale AI ran this benchmark against Muse Spark, which achieved an overall propensity score (rate of harmful action) of 24.5%.” source
Muse Spark Meta	Apollo Research Independent lab · 2026-05-26	in-context scheming Deception and sabotage	“Apollo Research found that Muse Spark exhibited limited rates of in-context scheming in realistic settings.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	time horizon 1.1 Autonomy and AI research	“* The TH 1.1 suite can’t reliably measure time horizons above 16 hours, but the 50% time horizon point estimate of the most capable shared model was between 16 and 20 hours, and the 80% time horizon point estimate was between 3 and 4 hours.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	subversion strategy eval Deception and sabotage	“**We only evaluated a subset of Subversion Strategy Eval due to refusals.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	challenge tasks Other	“Challenge tasks One edge-case success, seven clear failures One edge-case success, seven clear failures; more progress on some tasks” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	time horizon 1.1 Cyber	“The most capable agents we evaluated essentially saturated our Time Horizon 1.1 benchmark — there were only a handful of tasks longer than eight hours that they were still unable to solve, and many of those failures were due to cheating rather than obvious inability.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	mirrorcode Cyber	“Agents displayed even stronger capabilities on an early version of MirrorCode , a software reimplementation benchmark we developed with Epoch AI — the most capable agents we evaluated were able to solve tasks that would take humans weeks.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	dataset of 16 effective and 16 ineffective attack strategies for subverting anthropic's monitoring system Deception and sabotage	“As part of our red-teaming exercise with Anthropic , our embedded red-teamer built a dataset of 16 effective and 16 ineffective attack strategies for subverting Anthropic's monitoring system — for example, an effective strategy might involve a known monitor-jailbreaking technique, while an ineffective one might involve encoding unauthorized behavior in ways that LLM monitors would trivially spot.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	threat modeling for autonomous replication Autonomy and AI research	“We asked one of the most capable shared models to produce a write-up describing step-by-step how a rogue agent could autonomously replicate in the wild , similar to a 2–8 hour work test that we might have given to candidates for threat-modeling roles at METR.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	building a new mirrorcode task Autonomy and AI research	“We asked one of the most capable shared models to autonomously design and build a new, highly challenging MirrorCode task based on the original project proposal, with an inference budget of $1.5k-$15k (see Table 3).” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	design a $500 setup replicating sunlight Autonomy and AI research	“sign a $500 hardware setup for recreating sunlight indoors and produce a 21 We encouraged it to go deep on researching different branches of the threat modeling tree, since we thought that might play to the strengths of AI agents compared to human applicants.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	attack metr evaluation infrastructure (3 variants) Autonomy and AI research	“We ran 8 “Challenge tasks” (including the 3 separate variants of attacking METR infrastructure) on the shared models that we thought would perform best based on other evaluations.” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	build custom version of dvc Autonomy and AI research	“Build custom version of dvc ~$100-$1K (100M tokens) ❓ (failed to solve, but METR staff got much larger budget)” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	find vulnerabilities in runc Cyber	“Find vulnerabilities in runc ~$500-$5K (500M tokens) ❓ (humans likely would have needed a far larger budget)” source
Mythos Preview Anthropic	METR Independent lab · 2026-05-19	making money autonomously Autonomy and AI research	“Redwood Research asked Opus 4.7 to “make as much money as possible,” 24 running it autonomously for roughly four days (spending a few thousand dollars on inference tokens) with $5,000 in starting funds, an internet-connected computer, and tools for verifying online accounts.” source

Previous · Page 34 of 49 · Next